Genuinely trying to understand.
https://arxiv.org/abs/2206.15475 This paper summarizes it well
i enjoy implementing random ideas on a very badly coded version of the game pong.
Check the work that Judea Pearl won the Turing prize for.
I read his book, I understand and work with causal inference. I still don’t have an ELI5 understanding of what Causal ML is.
We do some work on causal discovery from observational data. For an overview of this line of work you can have a look at this survey paper D’ya like DAGs? A Survey on Structure Learning and Causal Discovery.
As someone who loves snatch and works with DAGs every day this is an excellent title
causal inference, leveraging machine learning methods to either estimate causal effects, infer causal graphs, or both. if you’re not familiar with “causal inference”, start with learning about that first and then move up to ml applications within that domain after getting your feet wet with CI first. Judea Pearl’s “The Book of Why” is a good introduction to the topic.
Thank you! I’m comfortable with causal inference and have read “The Book of Why”. What do you recommend for the next step?
One place might be to look at Granger Causality. I believe that causal ML can look for patterns in data that appear to conform to granger causality structures (ie there’s a leading and lagging indicator, if one always tracks the other then we can start to consider causality).
Normally causality is established in an experiment or natural experiment where we can isolate factors but since we have so much transactional data we can start to see patterns that resemble these structures without delineating the natural experiment ahead of time.
But causality is often very hard outside of very careful structures and it’s still a very active area.
One related place to look is also at network models like SAOMs which use panel data to explore issues with selection versus influence.
This is something I’ve been wondering as well. Is causal ML or causal analysis used in industry?
It’s situational, but yes.
Imagine you offer two membership tiers. You notice that people in the higher tier spend much more money with you. Yay!
Question: is the additional spend caused by (the fringe benefits of) the higher tier? If so, if you gave people a complementary upgrade, they’d spend more money with you. Win win. Or maybe people with more money to spend naturally tend to go for the higher tier, in which case your intervention will come to naught.
In this case, it’s easy to run an A/B test and find out. But in a lot of cases applying this kind of intervention can be difficult (because of cost, signal delay, amount of additional confounders) or downright immoral.
I have been wondering about this too. Never saw a single situation where causal modeling was practically applied.
I am thinking whether potentially there is a fundamental flaw here.
If you don’t know what causes what but just observe the correlation then most likely you will never find out what is the underlying cause because the entire causation is so complicated or hard to find or lacking further data that it’s just not feasible to figure it out.
If you know what causes what you can build your model accordingly. But in such a situation, the causality is usually not “perfect”. For example: If A causes B every single time when B occurs, then B does not truly give you any further information at all, because apparently all information must be contained in A. If, however, B does occur only with a certain probability when A occurs then you again are back to not knowing exactly how causality works, and there are unkonwn factors you cannot account for.
Don’t know, just some thoughts.
My backgrpund is in ux design and strategic planning and I’ve recently started learning about Causal ML (as part of learning ML).
I found that you already need to bring your assumption/reasoning of causation to the model as a flow chart, it’s usually more than A leads to B. Causal ML then uses your data to predict the accuracy of your assumption.
Here’s an obvious example:
If I have an A/B test of an ecommerce checkout, one blue and one red button. and in an a/b testing, red performs better. then a prediction model would learn that red performs better than blue.
In Causal ML, i would bring all factors in: background color, position, button color font, user & purchase information.
I can then create a first causal discovery model to come up with a network graph of the relation and then use causal ML to calculate the probability of effects.
Turns out, the color mages a difference for older shoppers because blue has lower contrast than red. so I could also choose another color with similar contrast for the same perfomance effect.
But I agree with you, a lot of that the thought work needs to be done before.
You know how ppl say correlation doesn’t imply causation? Well it does under some assumptions/conditions. Coming up with these conditions for specific datasets so that you can conclude correlation is causation is what it’s all about
You cannot find the causality without the agent manipulating the environment. Getting causality from observational data alone is impossible.
You can according to Pearl :)
I don’t think I’ve seen pearl say something like that
https://stats.stackexchange.com/questions/384330/is-causal-inference-only-from-data-possible. Here’s a post about why casual effects from data alone are not possible
Very interesting statistical theory, with very limited application.
I think the idea behind machine learning is that it can learn correlation and causality automatically. Can someone give an example of using causal ML to better solve a problem than non-causal ML methods?
You framed my question in a better way, thank you.
Casual ML has to do with t-shirts, cargo shorts, etc.
They’re probably working on causal inference. When you mention causal inference, I naturally think of causal graphs and linear models (and maybe occasionally random forests), so maybe that’s where people get the distinction? One thing in this domain I’ve worked on (in medium-sized tech) is notifications:
We say that we want to send exactly x notifications per user per day. Then train a model to predict P(DAU | send k notifications that day) and send the notifications that give you the highest P(DAU) uplift.
Some people would probably call this Causal ML; I didn’t think about confounders or causal graphs a single time while working on this, so I wouldn’t say I was working on causal inference here (I’d just say I was doing ML, but hmm maybe I should update my resume to say “Causal ML”…)
Causality is linked to disentanglement and sparse solutions. If you assume that there is a true causal representation, then there are ways to provably recover any permutation of such a representation given some assumptions. And being causal, the representation will also naturally be disentangled and sparse (a cat will clearly be separated from a dog).
See https://proceedings.mlr.press/v202/lachapelle23a/lachapelle23a.pdf.
I will say I am a little bit skeptical of this field but here is an example: https://arxiv.org/pdf/2306.05415.pdf
And a survey: